Forecast accuracy: the KPI most mid-sized companies never measure

Ask a room of managers what their forecast accuracy was last quarter, and you will usually get a thoughtful silence. The company forecasts constantly — sales does it, purchasing does it, the budget does it — but nobody writes those numbers down, freezes them, and checks three months later which one was actually right.

That is remarkable, because the forecast quietly steers the most expensive decisions in the business: what to buy, what to build, how much cash to tie up in stock. And a forecast error never just disappears. It turns into expedited freight, overtime, write-downs on stock nobody needed — or into empty shelves and customers who quietly order somewhere else.

The good news: measuring forecast quality is neither expensive nor complicated. You need three numbers and the discipline to look at them every month.

1. Bias: are you systematically wrong in one direction?

Bias is the simplest and most valuable of the three. Take all forecasts for a period, compare them with actual demand, and look at the sign of the error. A healthy forecast misses high about as often as it misses low. A biased forecast misses in the same direction month after month.

Bias almost always has an organizational cause, not a statistical one. Sales forecasts high because optimism is rewarded and pipeline pressure is real. Operations forecasts low because nobody wants to be the one who over-ordered. Neither side is lying — they are responding to incentives.

Here is the useful distinction: a forecast that is randomly wrong is a statistics problem. A forecast that is always 15% too high is a process problem — and process problems are fixable once they are visible.

2. Weighted MAPE: how big is the miss?

MAPE — mean absolute percentage error — answers “by how many percent are we off, on average?” An accuracy of 92% (a MAPE of 8%) means your forecasts miss actual demand by 8% on average, in either direction.

Two practical refinements matter:

• Weight it by revenue or margin. A 40% miss on a C-item that sells twice a year should not count the same as a 10% miss on your top seller. Weighted MAPE (WMAPE) keeps the metric honest.
• Measure at the level where decisions are made. Purchasing decides per article, so measure per article. Capacity decisions happen per product family, so also track the family level. Aggregate accuracy always looks flattering, because errors cancel out — that is exactly why it hides problems.

3. Forecast value added: is the effort worth it?

This one stings, which is why it is so useful. Compare your forecasting process against the dumbest possible benchmark: “next month equals the average of the last three months”, or for seasonal businesses, “same as last year”. That is called a naive forecast, and it costs nothing to produce.

Forecast value added (FVA) asks: does each step of your process — the statistical model, the sales override, the management adjustment — actually beat the naive forecast? In many companies, large planning meetings and manual overrides make the forecast worse than the naive baseline. If that is the case, you have just discovered free time and better forecasts in one analysis. In enterprise planning organizations this humility check is standard practice; it deserves to be in the mid-sized world too.

How to start, in Excel, this month

1. Freeze the forecast. At each month end, save the current forecast as plain values into an archive sheet. Never overwrite history — the single most common reason companies cannot measure accuracy is that the forecast is “alive” and yesterday’s numbers are gone.
2. Pick the lag that matters. Compare actuals against the forecast you had one lead time ago. If you reorder with a three-month horizon, the relevant question is what you believed three months ago.
3. Build one pivot table. Bias and WMAPE by product family, trended over time. That is genuinely all the math you need to begin.
4. Review it monthly, for 30 minutes. Same three charts every month. Discuss the five biggest misses — not all of them — and write down one action each.

What good looks like

Forget industry benchmark tables; accuracy depends too heavily on your product mix, granularity and horizon to compare across companies. What matters is the trend: bias moving toward zero, WMAPE drifting down quarter by quarter, overrides that demonstrably add value.

Takeaway — You do not need software to start measuring forecast accuracy. Freeze every forecast, compare at your lead-time lag, track bias and weighted MAPE by family, and review the numbers monthly. Companies that do this consistently get better forecasts within two or three cycles — not because the math is clever, but because forecasts suddenly have owners.